Accelerated Diagnosis of BRCA1/2-Related Breast Cancer Using Quantitative Computational Analysis of Consecutive DCE-MRI Scans, an AI Based Feasibility Study
2Oncology Institute, Sheba Medical Center Tel-Hashomer, Israel
Purpose: BRCA 1/2 carriers are at high-risk for developing breast cancer (BC), likely to be high-grade, aggressive and with an early-age onset. Carriers undergo an intensive surveillance, including annual MRI. This helps in reducing advanced-stage diagnosis. However, a visual retrospective inspection revealed that in ~60% of cases (total of 42 patients) enhancement was present in the tumor area of the prior scan (~1 year before diagnosis). We therefore aim at capturing that early enhancement using computational analysis of DCE-MRI (Dynamic Contrast Enhanced MRI) scans to achieve an even earlier diagnosis.
Hypothesis: Quantitative computational analysis of DCE-MRI with AI (Artificial Intelligence) techniques may identify the early enhancement and provide an “Earlier than Early” diagnosis of BRCA1/2-related BC.
Methods: A retrospective cohort of 164 breast DCE-MRIs (82 subjects). Group 1: diagnosis and preceding scan of 42 BC patients. Group 2: 40 pairs of consecutive cancer-free scans. Tumors were manually delineated and a set of ~250 features were extracted from the corresponding regions in the preceding scans (~1 year before). These included both “deep radiomics” and “curveology” features, from which a selection of 10 features was performed using predictive and information-theoretic criteria (Area Under ROC Curve, Jansen-Shannon divergence). Performance was assessed in a series of cross-validation experiments, each training and testing multiple SVM (Support Vector Machines) models.
Results: A 10-feature set successfully identified BC in 34/42 (~80%) cases, based on the scans from ~1 year before diagnosis. Future work will include, among other things, the reduction of false-positive rates (~33%).
Conclusions: The presented method uses quantitative features & ML (Machine Learning) to successfully flag radiologically suspicious regions, even before they are visually diagnostic. Preliminary results support the feasibility of significantly accelerating the diagnosis of BRCA-related BC using the proposed algorihtm.